From c8cf8bfdb634891d43e3fc877a00afefc8d94218 Mon Sep 17 00:00:00 2001
From: Erosika <eri@plasticlabs.ai>
Date: Thu, 6 Aug 2026 14:20:14 -0400
Subject: [PATCH] fix(agent): strip length-continuation marks from outgoing api
 messages

the scaffolding marks are hermes bookkeeping. only the chat-completions
transport strips underscore keys, so anthropic and bedrock requests on
continuation attempts 2+ would send the marks to strict providers. pop
them in the central api_messages sanitization next to _thinking_prefill.

also pin that a mark reloaded from a mid-crash persist on a prior turn's
message is never deleted by a later turn's ceiling cleanup.
---
 agent/conversation_loop.py                    |  3 +
 .../test_continuation_ceiling_wedge.py        | 63 +++++++++++++++++++
 2 files changed, 66 insertions(+)

diff --git a/agent/conversation_loop.py b/agent/conversation_loop.py
index 0603594c63487..d0e76d05dfce1 100644
--- a/agent/conversation_loop.py
+++ b/agent/conversation_loop.py
@@ -1798,6 +1798,9 @@ def run_conversation(
                 api_msg.pop("finish_reason")
             # Strip internal thinking-prefill marker
             api_msg.pop("_thinking_prefill", None)
+            # Strip length-continuation marks; not every transport drops underscore keys.
+            api_msg.pop("_length_continuation_fragment", None)
+            api_msg.pop("_length_continuation_nudge", None)
             # Strip Codex Responses API fields (call_id, response_item_id) for
             # strict providers like Mistral, Fireworks, etc. that reject unknown fields.
             # Uses new dicts so the internal messages list retains the fields
diff --git a/tests/run_agent/test_continuation_ceiling_wedge.py b/tests/run_agent/test_continuation_ceiling_wedge.py
index bba451b1e38aa..96fd63e945142 100644
--- a/tests/run_agent/test_continuation_ceiling_wedge.py
+++ b/tests/run_agent/test_continuation_ceiling_wedge.py
@@ -148,6 +148,69 @@ class TestContinuationCeilingWedge:
             "The user-facing message must name the truncation."
         )
 
+    def test_continuation_requests_carry_no_marks(self, loop_agent):
+        """The scaffolding marks are Hermes bookkeeping. The centrally
+        sanitized api_messages must never carry them — only the
+        chat-completions transport strips underscore keys, so anthropic
+        and bedrock requests would otherwise send them to the provider."""
+        from tests.run_agent.test_run_agent import _mock_response
+
+        seen_api_messages = []
+        original = loop_agent._build_api_kwargs
+
+        def _spy(api_messages, tools_for_api=None):
+            seen_api_messages.append([dict(m) for m in api_messages if isinstance(m, dict)])
+            return original(api_messages, tools_for_api=tools_for_api)
+
+        loop_agent.client.chat.completions.create.side_effect = [
+            _stub("part one "), _stub("part two "),
+            _mock_response(content="the rest.", finish_reason="stop"),
+        ]
+        with patch.object(loop_agent, "_build_api_kwargs", side_effect=_spy):
+            result = _run(loop_agent, "write me a long report")
+
+        assert result["completed"] is True
+        assert len(seen_api_messages) >= 3, "Expected continuation attempts 2+."
+        marked = [
+            (idx, key)
+            for idx, batch in enumerate(seen_api_messages)
+            for m in batch
+            for key in m
+            if str(key).startswith("_length_continuation")
+        ]
+        assert marked == [], (
+            f"Continuation marks leaked into outgoing api_messages: {marked!r}"
+        )
+
+    def test_prior_turn_marked_message_survives_later_ceiling(self, loop_agent):
+        """A mark that reached disk mid-crash and got reloaded on a PRIOR
+        turn's message must never be deleted by a later turn's ceiling
+        cleanup — the cleanup is scoped to the current turn."""
+        reloaded_history = [
+            {"role": "user", "content": "earlier question"},
+            {
+                "role": "assistant",
+                "content": "earlier answer fragment",
+                "_length_continuation_fragment": True,
+            },
+        ]
+        loop_agent.client.chat.completions.create.side_effect = [
+            _stub("wedge one "), _stub("wedge two "),
+            _stub("wedge three "), _stub("wedge four."),
+        ]
+        result = _run(loop_agent, "another long report", history=reloaded_history)
+
+        assert "truncated after 4 continuation attempts" in (result.get("error") or "")
+        prior = [
+            m for m in result["messages"]
+            if m.get("role") == "assistant"
+            and "earlier answer fragment" in (m.get("content") or "")
+        ]
+        assert len(prior) == 1, (
+            "The prior turn's reloaded message must survive the later "
+            "turn's ceiling cleanup."
+        )
+
     def test_new_turn_does_not_inherit_continuation_counter(self, loop_agent):
         """A single truncation on the turn AFTER the ceiling must get its
         own full 4-attempt budget, not the exhausted counter."""